4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; if not, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
26 * Copyright (c) 2012, 2014, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Author: liang@whamcloud.com
35 #define DEBUG_SUBSYSTEM S_LNET
37 #include <linux/cpu.h>
38 #include <linux/sched.h>
39 #include <libcfs/libcfs.h>
44 * modparam for setting number of partitions
46 * 0 : estimate best value based on cores or NUMA nodes
47 * 1 : disable multiple partitions
48 * >1 : specify number of partitions
50 static int cpu_npartitions;
51 CFS_MODULE_PARM(cpu_npartitions, "i", int, 0444, "# of CPU partitions");
54 * modparam for setting CPU partitions patterns:
56 * i.e: "0[0,1,2,3] 1[4,5,6,7]", number before bracket is CPU partition ID,
57 * number in bracket is processor ID (core or HT)
59 * i.e: "N 0[0,1] 1[2,3]" the first character 'N' means numbers in bracket
60 * are NUMA node ID, number before bracket is CPU partition ID.
62 * i.e: "N", shortcut expression to create CPT from NUMA & CPU topology
64 * NB: If user specified cpu_pattern, cpu_npartitions will be ignored
66 static char *cpu_pattern = "";
67 CFS_MODULE_PARM(cpu_pattern, "s", charp, 0444, "CPU partitions pattern");
70 /* serialize hotplug etc */
72 /* reserved for hotplug */
73 unsigned long cpt_version;
74 /* mutex to protect cpt_cpumask */
75 struct mutex cpt_mutex;
76 /* scratch buffer for set/unset_node */
77 cpumask_t *cpt_cpumask;
80 static struct cfs_cpt_data cpt_data;
83 cfs_cpu_core_siblings(int cpu, cpumask_t *mask)
85 /* return cpumask of cores in the same socket */
86 cpumask_copy(mask, topology_core_cpumask(cpu));
88 EXPORT_SYMBOL(cfs_cpu_core_siblings);
90 /* return number of cores in the same socket of \a cpu */
92 cfs_cpu_core_nsiblings(int cpu)
96 mutex_lock(&cpt_data.cpt_mutex);
98 cfs_cpu_core_siblings(cpu, cpt_data.cpt_cpumask);
99 num = cpus_weight(*cpt_data.cpt_cpumask);
101 mutex_unlock(&cpt_data.cpt_mutex);
105 EXPORT_SYMBOL(cfs_cpu_core_nsiblings);
107 /* return cpumask of HTs in the same core */
109 cfs_cpu_ht_siblings(int cpu, cpumask_t *mask)
111 cpumask_copy(mask, topology_thread_cpumask(cpu));
113 EXPORT_SYMBOL(cfs_cpu_ht_siblings);
115 /* return number of HTs in the same core of \a cpu */
117 cfs_cpu_ht_nsiblings(int cpu)
121 mutex_lock(&cpt_data.cpt_mutex);
123 cfs_cpu_ht_siblings(cpu, cpt_data.cpt_cpumask);
124 num = cpus_weight(*cpt_data.cpt_cpumask);
126 mutex_unlock(&cpt_data.cpt_mutex);
130 EXPORT_SYMBOL(cfs_cpu_ht_nsiblings);
133 cfs_node_to_cpumask(int node, cpumask_t *mask)
135 const cpumask_t *tmp = cpumask_of_node(node);
138 cpumask_copy(mask, tmp);
142 EXPORT_SYMBOL(cfs_node_to_cpumask);
145 cfs_cpt_table_free(struct cfs_cpt_table *cptab)
149 if (cptab->ctb_cpu2cpt != NULL) {
150 LIBCFS_FREE(cptab->ctb_cpu2cpt,
151 num_possible_cpus() *
152 sizeof(cptab->ctb_cpu2cpt[0]));
155 for (i = 0; cptab->ctb_parts != NULL && i < cptab->ctb_nparts; i++) {
156 struct cfs_cpu_partition *part = &cptab->ctb_parts[i];
158 if (part->cpt_nodemask != NULL) {
159 LIBCFS_FREE(part->cpt_nodemask,
160 sizeof(*part->cpt_nodemask));
163 if (part->cpt_cpumask != NULL)
164 LIBCFS_FREE(part->cpt_cpumask, cpumask_size());
167 if (cptab->ctb_parts != NULL) {
168 LIBCFS_FREE(cptab->ctb_parts,
169 cptab->ctb_nparts * sizeof(cptab->ctb_parts[0]));
172 if (cptab->ctb_nodemask != NULL)
173 LIBCFS_FREE(cptab->ctb_nodemask, sizeof(*cptab->ctb_nodemask));
174 if (cptab->ctb_cpumask != NULL)
175 LIBCFS_FREE(cptab->ctb_cpumask, cpumask_size());
177 LIBCFS_FREE(cptab, sizeof(*cptab));
179 EXPORT_SYMBOL(cfs_cpt_table_free);
181 struct cfs_cpt_table *
182 cfs_cpt_table_alloc(unsigned int ncpt)
184 struct cfs_cpt_table *cptab;
187 LIBCFS_ALLOC(cptab, sizeof(*cptab));
191 cptab->ctb_nparts = ncpt;
193 LIBCFS_ALLOC(cptab->ctb_cpumask, cpumask_size());
194 LIBCFS_ALLOC(cptab->ctb_nodemask, sizeof(*cptab->ctb_nodemask));
196 if (cptab->ctb_cpumask == NULL || cptab->ctb_nodemask == NULL)
199 LIBCFS_ALLOC(cptab->ctb_cpu2cpt,
200 num_possible_cpus() * sizeof(cptab->ctb_cpu2cpt[0]));
201 if (cptab->ctb_cpu2cpt == NULL)
204 memset(cptab->ctb_cpu2cpt, -1,
205 num_possible_cpus() * sizeof(cptab->ctb_cpu2cpt[0]));
207 LIBCFS_ALLOC(cptab->ctb_parts, ncpt * sizeof(cptab->ctb_parts[0]));
208 if (cptab->ctb_parts == NULL)
211 for (i = 0; i < ncpt; i++) {
212 struct cfs_cpu_partition *part = &cptab->ctb_parts[i];
214 LIBCFS_ALLOC(part->cpt_cpumask, cpumask_size());
215 LIBCFS_ALLOC(part->cpt_nodemask, sizeof(*part->cpt_nodemask));
216 if (part->cpt_cpumask == NULL || part->cpt_nodemask == NULL)
220 spin_lock(&cpt_data.cpt_lock);
221 /* Reserved for hotplug */
222 cptab->ctb_version = cpt_data.cpt_version;
223 spin_unlock(&cpt_data.cpt_lock);
228 cfs_cpt_table_free(cptab);
231 EXPORT_SYMBOL(cfs_cpt_table_alloc);
234 cfs_cpt_table_print(struct cfs_cpt_table *cptab, char *buf, int len)
241 for (i = 0; i < cptab->ctb_nparts; i++) {
243 rc = snprintf(tmp, len, "%d\t: ", i);
253 for_each_cpu_mask(j, *cptab->ctb_parts[i].cpt_cpumask) {
254 rc = snprintf(tmp, len, "%d ", j);
274 EXPORT_SYMBOL(cfs_cpt_table_print);
277 cfs_cpt_number(struct cfs_cpt_table *cptab)
279 return cptab->ctb_nparts;
281 EXPORT_SYMBOL(cfs_cpt_number);
284 cfs_cpt_weight(struct cfs_cpt_table *cptab, int cpt)
286 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
288 return cpt == CFS_CPT_ANY ?
289 cpus_weight(*cptab->ctb_cpumask) :
290 cpus_weight(*cptab->ctb_parts[cpt].cpt_cpumask);
292 EXPORT_SYMBOL(cfs_cpt_weight);
295 cfs_cpt_online(struct cfs_cpt_table *cptab, int cpt)
297 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
299 return cpt == CFS_CPT_ANY ?
300 any_online_cpu(*cptab->ctb_cpumask) != NR_CPUS :
301 any_online_cpu(*cptab->ctb_parts[cpt].cpt_cpumask) != NR_CPUS;
303 EXPORT_SYMBOL(cfs_cpt_online);
306 cfs_cpt_cpumask(struct cfs_cpt_table *cptab, int cpt)
308 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
310 return cpt == CFS_CPT_ANY ?
311 cptab->ctb_cpumask : cptab->ctb_parts[cpt].cpt_cpumask;
313 EXPORT_SYMBOL(cfs_cpt_cpumask);
316 cfs_cpt_nodemask(struct cfs_cpt_table *cptab, int cpt)
318 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
320 return cpt == CFS_CPT_ANY ?
321 cptab->ctb_nodemask : cptab->ctb_parts[cpt].cpt_nodemask;
323 EXPORT_SYMBOL(cfs_cpt_nodemask);
326 cfs_cpt_set_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu)
330 LASSERT(cpt >= 0 && cpt < cptab->ctb_nparts);
332 if (cpu < 0 || cpu >= NR_CPUS || !cpu_online(cpu)) {
333 CDEBUG(D_INFO, "CPU %d is invalid or it's offline\n", cpu);
337 if (cptab->ctb_cpu2cpt[cpu] != -1) {
338 CDEBUG(D_INFO, "CPU %d is already in partition %d\n",
339 cpu, cptab->ctb_cpu2cpt[cpu]);
343 cptab->ctb_cpu2cpt[cpu] = cpt;
345 LASSERT(!cpu_isset(cpu, *cptab->ctb_cpumask));
346 LASSERT(!cpu_isset(cpu, *cptab->ctb_parts[cpt].cpt_cpumask));
348 cpu_set(cpu, *cptab->ctb_cpumask);
349 cpu_set(cpu, *cptab->ctb_parts[cpt].cpt_cpumask);
351 node = cpu_to_node(cpu);
353 /* first CPU of @node in this CPT table */
354 if (!node_isset(node, *cptab->ctb_nodemask))
355 node_set(node, *cptab->ctb_nodemask);
357 /* first CPU of @node in this partition */
358 if (!node_isset(node, *cptab->ctb_parts[cpt].cpt_nodemask))
359 node_set(node, *cptab->ctb_parts[cpt].cpt_nodemask);
363 EXPORT_SYMBOL(cfs_cpt_set_cpu);
366 cfs_cpt_unset_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu)
371 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
373 if (cpu < 0 || cpu >= NR_CPUS) {
374 CDEBUG(D_INFO, "Invalid CPU id %d\n", cpu);
378 if (cpt == CFS_CPT_ANY) {
379 /* caller doesn't know the partition ID */
380 cpt = cptab->ctb_cpu2cpt[cpu];
381 if (cpt < 0) { /* not set in this CPT-table */
382 CDEBUG(D_INFO, "Try to unset cpu %d which is "
383 "not in CPT-table %p\n", cpt, cptab);
387 } else if (cpt != cptab->ctb_cpu2cpt[cpu]) {
389 "CPU %d is not in cpu-partition %d\n", cpu, cpt);
393 LASSERT(cpu_isset(cpu, *cptab->ctb_parts[cpt].cpt_cpumask));
394 LASSERT(cpu_isset(cpu, *cptab->ctb_cpumask));
396 cpu_clear(cpu, *cptab->ctb_parts[cpt].cpt_cpumask);
397 cpu_clear(cpu, *cptab->ctb_cpumask);
398 cptab->ctb_cpu2cpt[cpu] = -1;
400 node = cpu_to_node(cpu);
402 LASSERT(node_isset(node, *cptab->ctb_parts[cpt].cpt_nodemask));
403 LASSERT(node_isset(node, *cptab->ctb_nodemask));
405 for_each_cpu_mask(i, *cptab->ctb_parts[cpt].cpt_cpumask) {
406 /* this CPT has other CPU belonging to this node? */
407 if (cpu_to_node(i) == node)
412 node_clear(node, *cptab->ctb_parts[cpt].cpt_nodemask);
414 for_each_cpu_mask(i, *cptab->ctb_cpumask) {
415 /* this CPT-table has other CPU belonging to this node? */
416 if (cpu_to_node(i) == node)
421 node_clear(node, *cptab->ctb_nodemask);
425 EXPORT_SYMBOL(cfs_cpt_unset_cpu);
428 cfs_cpt_set_cpumask(struct cfs_cpt_table *cptab, int cpt, cpumask_t *mask)
432 if (cpus_weight(*mask) == 0 || any_online_cpu(*mask) == NR_CPUS) {
433 CDEBUG(D_INFO, "No online CPU is found in the CPU mask "
434 "for CPU partition %d\n", cpt);
438 for_each_cpu_mask(i, *mask) {
439 if (!cfs_cpt_set_cpu(cptab, cpt, i))
445 EXPORT_SYMBOL(cfs_cpt_set_cpumask);
448 cfs_cpt_unset_cpumask(struct cfs_cpt_table *cptab, int cpt, cpumask_t *mask)
452 for_each_cpu_mask(i, *mask)
453 cfs_cpt_unset_cpu(cptab, cpt, i);
455 EXPORT_SYMBOL(cfs_cpt_unset_cpumask);
458 cfs_cpt_set_node(struct cfs_cpt_table *cptab, int cpt, int node)
463 if (node < 0 || node >= MAX_NUMNODES) {
465 "Invalid NUMA id %d for CPU partition %d\n", node, cpt);
469 mutex_lock(&cpt_data.cpt_mutex);
471 mask = cpt_data.cpt_cpumask;
472 cfs_node_to_cpumask(node, mask);
474 rc = cfs_cpt_set_cpumask(cptab, cpt, mask);
476 mutex_unlock(&cpt_data.cpt_mutex);
480 EXPORT_SYMBOL(cfs_cpt_set_node);
483 cfs_cpt_unset_node(struct cfs_cpt_table *cptab, int cpt, int node)
487 if (node < 0 || node >= MAX_NUMNODES) {
489 "Invalid NUMA id %d for CPU partition %d\n", node, cpt);
493 mutex_lock(&cpt_data.cpt_mutex);
495 mask = cpt_data.cpt_cpumask;
496 cfs_node_to_cpumask(node, mask);
498 cfs_cpt_unset_cpumask(cptab, cpt, mask);
500 mutex_unlock(&cpt_data.cpt_mutex);
502 EXPORT_SYMBOL(cfs_cpt_unset_node);
505 cfs_cpt_set_nodemask(struct cfs_cpt_table *cptab, int cpt, nodemask_t *mask)
509 for_each_node_mask(i, *mask) {
510 if (!cfs_cpt_set_node(cptab, cpt, i))
516 EXPORT_SYMBOL(cfs_cpt_set_nodemask);
519 cfs_cpt_unset_nodemask(struct cfs_cpt_table *cptab, int cpt, nodemask_t *mask)
523 for_each_node_mask(i, *mask)
524 cfs_cpt_unset_node(cptab, cpt, i);
526 EXPORT_SYMBOL(cfs_cpt_unset_nodemask);
529 cfs_cpt_clear(struct cfs_cpt_table *cptab, int cpt)
534 if (cpt == CFS_CPT_ANY) {
535 last = cptab->ctb_nparts - 1;
541 for (; cpt <= last; cpt++) {
542 for_each_cpu_mask(i, *cptab->ctb_parts[cpt].cpt_cpumask)
543 cfs_cpt_unset_cpu(cptab, cpt, i);
546 EXPORT_SYMBOL(cfs_cpt_clear);
549 cfs_cpt_spread_node(struct cfs_cpt_table *cptab, int cpt)
556 /* convert CPU partition ID to HW node id */
558 if (cpt < 0 || cpt >= cptab->ctb_nparts) {
559 mask = cptab->ctb_nodemask;
560 rotor = cptab->ctb_spread_rotor++;
562 mask = cptab->ctb_parts[cpt].cpt_nodemask;
563 rotor = cptab->ctb_parts[cpt].cpt_spread_rotor++;
566 weight = nodes_weight(*mask);
571 for_each_node_mask(node, *mask) {
579 EXPORT_SYMBOL(cfs_cpt_spread_node);
582 cfs_cpt_current(struct cfs_cpt_table *cptab, int remap)
584 int cpu = smp_processor_id();
585 int cpt = cptab->ctb_cpu2cpt[cpu];
591 /* don't return negative value for safety of upper layer,
592 * instead we shadow the unknown cpu to a valid partition ID */
593 cpt = cpu % cptab->ctb_nparts;
598 EXPORT_SYMBOL(cfs_cpt_current);
601 cfs_cpt_of_cpu(struct cfs_cpt_table *cptab, int cpu)
603 LASSERT(cpu >= 0 && cpu < NR_CPUS);
605 return cptab->ctb_cpu2cpt[cpu];
607 EXPORT_SYMBOL(cfs_cpt_of_cpu);
610 cfs_cpt_bind(struct cfs_cpt_table *cptab, int cpt)
613 nodemask_t *nodemask;
617 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
619 if (cpt == CFS_CPT_ANY) {
620 cpumask = cptab->ctb_cpumask;
621 nodemask = cptab->ctb_nodemask;
623 cpumask = cptab->ctb_parts[cpt].cpt_cpumask;
624 nodemask = cptab->ctb_parts[cpt].cpt_nodemask;
627 if (any_online_cpu(*cpumask) == NR_CPUS) {
628 CERROR("No online CPU found in CPU partition %d, did someone "
629 "do CPU hotplug on system? You might need to reload "
630 "Lustre modules to keep system working well.\n", cpt);
634 for_each_online_cpu(i) {
635 if (cpu_isset(i, *cpumask))
638 rc = set_cpus_allowed_ptr(current, cpumask);
639 set_mems_allowed(*nodemask);
641 schedule(); /* switch to allowed CPU */
646 /* don't need to set affinity because all online CPUs are covered */
649 EXPORT_SYMBOL(cfs_cpt_bind);
652 * Choose max to \a number CPUs from \a node and set them in \a cpt.
653 * We always prefer to choose CPU in the same core/socket.
656 cfs_cpt_choose_ncpus(struct cfs_cpt_table *cptab, int cpt,
657 cpumask_t *node, int number)
659 cpumask_t *socket = NULL;
660 cpumask_t *core = NULL;
666 if (number >= cpus_weight(*node)) {
667 while (!cpus_empty(*node)) {
668 cpu = first_cpu(*node);
670 rc = cfs_cpt_set_cpu(cptab, cpt, cpu);
673 cpu_clear(cpu, *node);
678 /* allocate scratch buffer */
679 LIBCFS_ALLOC(socket, cpumask_size());
680 LIBCFS_ALLOC(core, cpumask_size());
681 if (socket == NULL || core == NULL) {
686 while (!cpus_empty(*node)) {
687 cpu = first_cpu(*node);
689 /* get cpumask for cores in the same socket */
690 cfs_cpu_core_siblings(cpu, socket);
691 cpus_and(*socket, *socket, *node);
693 LASSERT(!cpus_empty(*socket));
695 while (!cpus_empty(*socket)) {
698 /* get cpumask for hts in the same core */
699 cfs_cpu_ht_siblings(cpu, core);
700 cpus_and(*core, *core, *node);
702 LASSERT(!cpus_empty(*core));
704 for_each_cpu_mask(i, *core) {
705 cpu_clear(i, *socket);
708 rc = cfs_cpt_set_cpu(cptab, cpt, i);
717 cpu = first_cpu(*socket);
723 LIBCFS_FREE(socket, cpumask_size());
725 LIBCFS_FREE(core, cpumask_size());
729 #define CPT_WEIGHT_MIN 4u
732 cfs_cpt_num_estimate(void)
734 unsigned nnode = num_online_nodes();
735 unsigned ncpu = num_online_cpus();
738 if (ncpu <= CPT_WEIGHT_MIN) {
743 /* generate reasonable number of CPU partitions based on total number
744 * of CPUs, Preferred N should be power2 and match this condition:
745 * 2 * (N - 1)^2 < NCPUS <= 2 * N^2 */
746 for (ncpt = 2; ncpu > 2 * ncpt * ncpt; ncpt <<= 1) {}
748 if (ncpt <= nnode) { /* fat numa system */
752 } else { /* ncpt > nnode */
753 while ((nnode << 1) <= ncpt)
760 #if (BITS_PER_LONG == 32)
761 /* config many CPU partitions on 32-bit system could consume
763 ncpt = min(2U, ncpt);
765 while (ncpu % ncpt != 0)
766 ncpt--; /* worst case is 1 */
771 static struct cfs_cpt_table *
772 cfs_cpt_table_create(int ncpt)
774 struct cfs_cpt_table *cptab = NULL;
775 cpumask_t *mask = NULL;
781 rc = cfs_cpt_num_estimate();
785 if (ncpt > num_online_cpus() || ncpt > 4 * rc) {
786 CWARN("CPU partition number %d is larger than suggested "
787 "value (%d), your system may have performance"
788 "issue or run out of memory while under pressure\n",
792 if (num_online_cpus() % ncpt != 0) {
793 CERROR("CPU number %d is not multiple of cpu_npartition %d, "
794 "please try different cpu_npartitions value or"
795 "set pattern string by cpu_pattern=STRING\n",
796 (int)num_online_cpus(), ncpt);
800 cptab = cfs_cpt_table_alloc(ncpt);
802 CERROR("Failed to allocate CPU map(%d)\n", ncpt);
806 num = num_online_cpus() / ncpt;
808 CERROR("CPU changed while setting CPU partition\n");
812 LIBCFS_ALLOC(mask, cpumask_size());
814 CERROR("Failed to allocate scratch cpumask\n");
818 for_each_online_node(i) {
819 cfs_node_to_cpumask(i, mask);
821 while (!cpus_empty(*mask)) {
822 struct cfs_cpu_partition *part;
825 /* Each emulated NUMA node has all allowed CPUs in
827 * End loop when all partitions have assigned CPUs.
832 part = &cptab->ctb_parts[cpt];
834 n = num - cpus_weight(*part->cpt_cpumask);
837 rc = cfs_cpt_choose_ncpus(cptab, cpt, mask, n);
841 LASSERT(num >= cpus_weight(*part->cpt_cpumask));
842 if (num == cpus_weight(*part->cpt_cpumask))
848 num != cpus_weight(*cptab->ctb_parts[ncpt - 1].cpt_cpumask)) {
849 CERROR("Expect %d(%d) CPU partitions but got %d(%d), "
850 "CPU hotplug/unplug while setting?\n",
851 cptab->ctb_nparts, num, cpt,
852 cpus_weight(*cptab->ctb_parts[ncpt - 1].cpt_cpumask));
856 LIBCFS_FREE(mask, cpumask_size());
861 CERROR("Failed to setup CPU-partition-table with %d "
862 "CPU-partitions, online HW nodes: %d, HW cpus: %d.\n",
863 ncpt, num_online_nodes(), num_online_cpus());
866 LIBCFS_FREE(mask, cpumask_size());
869 cfs_cpt_table_free(cptab);
874 static struct cfs_cpt_table *
875 cfs_cpt_table_create_pattern(char *pattern)
877 struct cfs_cpt_table *cptab;
887 str = cfs_trimwhite(pattern);
888 if (*str == 'n' || *str == 'N') {
890 if (*pattern != '\0') {
891 node = 1; /* numa pattern */
893 } else { /* shortcut to create CPT from NUMA & CPU topology */
895 ncpt = num_online_nodes();
899 if (ncpt == 0) { /* scanning bracket which is mark of partition */
900 for (str = pattern;; str++, ncpt++) {
901 str = strchr(str, '[');
908 (node && ncpt > num_online_nodes()) ||
909 (!node && ncpt > num_online_cpus())) {
910 CERROR("Invalid pattern %s, or too many partitions %d\n",
915 cptab = cfs_cpt_table_alloc(ncpt);
917 CERROR("Failed to allocate cpu partition table\n");
921 if (node < 0) { /* shortcut to create CPT from NUMA & CPU topology */
923 for_each_online_node(i) {
925 CERROR("CPU changed while setting CPU "
926 "partition table, %d/%d\n", cpt, ncpt);
930 rc = cfs_cpt_set_node(cptab, cpt++, i);
937 high = node ? MAX_NUMNODES - 1 : nr_cpu_ids - 1;
939 for (str = cfs_trimwhite(pattern), c = 0;; c++) {
940 struct cfs_range_expr *range;
941 struct cfs_expr_list *el;
942 char *bracket = strchr(str, '[');
945 if (bracket == NULL) {
947 CERROR("Invalid pattern %s\n", str);
949 } else if (c != ncpt) {
950 CERROR("expect %d partitions but found %d\n",
957 if (sscanf(str, "%d%n", &cpt, &n) < 1) {
958 CERROR("Invalid cpu pattern %s\n", str);
962 if (cpt < 0 || cpt >= ncpt) {
963 CERROR("Invalid partition id %d, total partitions %d\n",
968 if (cfs_cpt_weight(cptab, cpt) != 0) {
969 CERROR("Partition %d has already been set.\n", cpt);
973 str = cfs_trimwhite(str + n);
974 if (str != bracket) {
975 CERROR("Invalid pattern %s\n", str);
979 bracket = strchr(str, ']');
980 if (bracket == NULL) {
981 CERROR("missing right bracket for cpt %d, %s\n",
986 if (cfs_expr_list_parse(str, (bracket - str) + 1,
987 0, high, &el) != 0) {
988 CERROR("Can't parse number range: %s\n", str);
992 list_for_each_entry(range, &el->el_exprs, re_link) {
993 for (i = range->re_lo; i <= range->re_hi; i++) {
994 if ((i - range->re_lo) % range->re_stride != 0)
997 rc = node ? cfs_cpt_set_node(cptab, cpt, i) :
998 cfs_cpt_set_cpu(cptab, cpt, i);
1000 cfs_expr_list_free(el);
1006 cfs_expr_list_free(el);
1008 if (!cfs_cpt_online(cptab, cpt)) {
1009 CERROR("No online CPU is found on partition %d\n", cpt);
1013 str = cfs_trimwhite(bracket + 1);
1019 cfs_cpt_table_free(cptab);
1023 #ifdef CONFIG_HOTPLUG_CPU
1025 cfs_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
1027 unsigned int cpu = (unsigned long)hcpu;
1032 case CPU_DEAD_FROZEN:
1034 case CPU_ONLINE_FROZEN:
1035 spin_lock(&cpt_data.cpt_lock);
1036 cpt_data.cpt_version++;
1037 spin_unlock(&cpt_data.cpt_lock);
1039 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) {
1040 CDEBUG(D_INFO, "CPU changed [cpu %u action %lx]\n",
1045 mutex_lock(&cpt_data.cpt_mutex);
1046 /* if all HTs in a core are offline, it may break affinity */
1047 cfs_cpu_ht_siblings(cpu, cpt_data.cpt_cpumask);
1048 warn = any_online_cpu(*cpt_data.cpt_cpumask) >= nr_cpu_ids;
1049 mutex_unlock(&cpt_data.cpt_mutex);
1050 CDEBUG(warn ? D_WARNING : D_INFO,
1051 "Lustre: can't support CPU plug-out well now, "
1052 "performance and stability could be impacted"
1053 "[CPU %u action: %lx]\n", cpu, action);
1059 static struct notifier_block cfs_cpu_notifier = {
1060 .notifier_call = cfs_cpu_notify,
1069 if (cfs_cpt_table != NULL)
1070 cfs_cpt_table_free(cfs_cpt_table);
1072 #ifdef CONFIG_HOTPLUG_CPU
1073 unregister_hotcpu_notifier(&cfs_cpu_notifier);
1075 if (cpt_data.cpt_cpumask != NULL)
1076 LIBCFS_FREE(cpt_data.cpt_cpumask, cpumask_size());
1082 LASSERT(cfs_cpt_table == NULL);
1084 memset(&cpt_data, 0, sizeof(cpt_data));
1086 LIBCFS_ALLOC(cpt_data.cpt_cpumask, cpumask_size());
1087 if (cpt_data.cpt_cpumask == NULL) {
1088 CERROR("Failed to allocate scratch buffer\n");
1092 spin_lock_init(&cpt_data.cpt_lock);
1093 mutex_init(&cpt_data.cpt_mutex);
1095 #ifdef CONFIG_HOTPLUG_CPU
1096 register_hotcpu_notifier(&cfs_cpu_notifier);
1099 if (*cpu_pattern != 0) {
1100 cfs_cpt_table = cfs_cpt_table_create_pattern(cpu_pattern);
1101 if (cfs_cpt_table == NULL) {
1102 CERROR("Failed to create cptab from pattern %s\n",
1108 cfs_cpt_table = cfs_cpt_table_create(cpu_npartitions);
1109 if (cfs_cpt_table == NULL) {
1110 CERROR("Failed to create ptable with npartitions %d\n",
1116 spin_lock(&cpt_data.cpt_lock);
1117 if (cfs_cpt_table->ctb_version != cpt_data.cpt_version) {
1118 spin_unlock(&cpt_data.cpt_lock);
1119 CERROR("CPU hotplug/unplug during setup\n");
1122 spin_unlock(&cpt_data.cpt_lock);
1124 LCONSOLE(0, "HW CPU cores: %d, npartitions: %d\n",
1125 num_online_cpus(), cfs_cpt_number(cfs_cpt_table));